STSP 6813

Stochastic Process
University of the Free State
2024
Dr. Niladri Chakraborty
(Lecture notes 2)

1. Nonhomogeneous Poisson process

This is a generalization of the homogeneous Poisson process on the case of the time dependent
rate 𝜆(𝑡).
Definition: The counting (point) process {𝑁(𝑡), 𝑡 ≥ 0} is said to be a nonstationary (or
nonhomogeneous) Poisson process with rate 𝜆(𝑡), if,
i. 𝑁(0) = 0;
ii. {𝑁(𝑡), 𝑡 ≥ 0} has independent (but not stationary) increments
iii. 𝑃[𝑁(𝑡 + ℎ) − 𝑁(𝑡) ≥ 2] = 𝑜(ℎ)
iv. 𝑃[𝑁(𝑡 + ℎ) − 𝑁(𝑡) = 1] = 𝜆(𝑡)ℎ + 𝑜(ℎ)

𝑡
We know the definition of the cumulated rate as Λ(𝑡) = ∫0 𝜆(𝑢)𝑑𝑢

Result 1: It can be shown that

𝑡+𝑠 𝑛
𝜆(𝑢)𝑑𝑢 (∫𝑠 𝜆(𝑢) 𝑑𝑢)
𝑡+𝑠
− ∫𝑠
𝑃𝑛 (𝑡) = 𝑃[𝑁(𝑡 + 𝑠) − 𝑁(𝑠) = 𝑛] = 𝑒 .
𝑛!

The proof is like the one for homogeneous case. Specifically, 𝑃0 (𝑠) = 𝑃[𝑁(𝑡 + 𝑠) − 𝑁(𝑠) =
𝑡+𝑠
0] = 𝑒 − ∫𝑠 𝜆(𝑢)𝑑𝑢
.

Example: (Minimal repair)

Consider the lifetime random variable 𝑋 ≥ 0 with a distribution function 𝑃[𝑋 ≤ 𝑡] = 𝐹(𝑡),
𝑓(𝑡)
and the failure rate 𝜆(𝑡) = 1−𝐹(𝑡). 𝜆(𝑡) means the probability of occurrence of failure in [𝑡, 𝑡 +
𝑑𝑡) given that it did not occur in [0, 𝑡).
Assume that the failure is minimally repaired that means that, after the first repair (instant) at
time 𝑡1 , the distribution function is
𝐹̅ (𝑡+𝑡1 ) 𝐹(𝑡+𝑡1 )−𝐹(𝑡1 )
𝐹̅ (𝑡|𝑡1 ) = 𝐹̅ (𝑡1 )
, or equivalently 𝐹(𝑡|𝑡1 ) = 𝐹(𝑡1 )
, similar to the second repair at 𝑡2 >
𝑡1 and etc.

1
Exercise: Show that the point process {𝑇𝑖 } of failures with instant minimal repair form a
nonhomogeneous Poisson process with rate 𝜆(𝑡).
Hint: Look at the interarrival times of the nonhomogeneous process.

1.1. Compound Poisson random variable:

Reminder:
The moment generating function (m.g.f.) for the random variable 𝑋 with a continuous
distribution 𝐹(𝑥), for 𝑡 ∈ [−1,1],
∞
𝜓(𝑡) = 𝐸[𝑒 𝑡𝑥 ] = ∫0 𝑒 𝑡𝑥 𝑑𝐹(𝑥).

For a discrete distribution. 𝜓(𝑡) = ∑∞

0 𝑒
𝑡𝑥
𝑃[𝑋 = 𝑥].

𝐸[𝑋 𝑛 ] = 𝜓 𝑛 (0), 𝑓𝑜𝑟 𝑛 ≥ 1.

𝑡 −1)
It can be shown that, for a Poisson distribution, 𝜓(𝑡) = 𝑒 𝜆(𝑒 .
Definition of a compound Poisson random variable is given by
𝑊 = ∑𝑁
𝑖=1 𝑋𝑖 ,

where {𝑋𝑖 } i.i.d. with 𝐹(𝑥), 𝑁 ∼ Poisson(𝜆), and 𝑁 is independent of 𝑋𝑖 .

The moment generating function of 𝑾:
𝑁
𝑡 ∑𝑖=1 𝑋𝑖
𝜓1 (𝑡) = 𝐸[𝑒 𝑡𝑊 ] = ∑∞ 𝑡𝑤 ∞
𝑛=0 𝐸[𝑒 |𝑁] 𝑃[𝑁 = 𝑛] = ∑𝑛=0 𝐸[𝑒 |𝑁 = 𝑛] 𝑃[𝑁 = 𝑛]
𝑁
Because 𝑋𝑖 ’s are independent, we can write, 𝐸 [𝑒 𝑡 ∑𝑖=1 𝑋𝑖 |𝑁 = 𝑛] = ∏𝑛𝑖=1 𝐸[𝑒 𝑡𝑥𝑖 ] = [𝜓(𝑡)]𝑛 .

Therefore,
𝑁
𝑡 ∑𝑖=1 𝑋𝑖
𝐸[𝑒 𝑡𝑊 ] = ∑∞
𝑛=0 𝐸[𝑒 |𝑁 = 𝑛] 𝑃[𝑁 = 𝑛]
𝑒 −𝜆𝑡 (𝜆𝑡)𝑛
= ∑∞ 𝑛 ∞
𝑛=0[𝜓(𝑡)] 𝑃[𝑁 = 𝑛] = ∑𝑛=0[𝜓(𝑡)]
𝑛
𝑛!

= 𝑒 𝜆𝑡(𝜓(𝑡)−1) .
𝑡 −1)
When 𝑋𝑖 ∼ Poisson(𝜆), we have 𝜓(𝑡) = 𝑒 𝜆(𝑒 (how ??).

By differentiating 𝜓1 (𝑡) and replacing 𝑡 = 0, we can get different moments of the distribution
of 𝑊. (How??)

2
 1.2. Conditional and compound Poisson distribution

Compound Poisson process:

Similar to the compound Poisson random variable, a compound Poisson process is defined as

𝑋(𝑡) = ∑𝑁(𝑡)
𝑖=1 𝑋𝑖

where {𝑁(𝑡); 𝑡 ≥ 0} is a Poisson process and {𝑋𝑖 } are i.i.d.

Conditional Poisson process:

Suppose Λ is a positive r.v. with distribution 𝐺(𝜆).
Let 𝑁(𝑡) be a counting process such that for every Λ = 𝜆, it is a Poisson process with rate 𝜆.
For example, the following property holds,
∞ 𝑒 −𝜆𝑡 (𝜆𝑡)𝑛
𝑃[𝑁(𝑡 + 𝑠) − 𝑁(𝑠) = 𝑛] = ∫0 𝑑𝐺(𝜆).
𝑛!

Strictly speaking, the conditional Poisson process is not a Poisson process because the
assumption of independent increment is violated.

There can be many problems connected to the conditional Poisson process. For example, we
may obtain the conditional distribution of Λ given 𝑁(𝑡) = 𝑛, when Λ ∼ 𝐺(𝜆).
𝑃[Λ ∈ [𝜆, 𝜆 + 𝑑𝜆)|𝑁(𝑡) = 𝑛]
𝑃[Λ∈[𝜆,𝜆+𝑑𝜆)]
= 𝑃[𝑁(𝑡) = 𝑛|Λ ∈ [𝜆, 𝜆 + 𝑑𝜆)]
𝑃[𝑁(𝑡)=𝑛]

𝑒 −𝜆𝑡 (𝜆𝑡)𝑛 𝑑𝐺(𝜆)

= ∞𝑒−𝜆𝑡 (𝜆𝑡) 𝑛
.
𝑛! ∫0 𝑑𝐺(𝜆)
𝑛!

Therefore,
𝑥𝑒−𝜆𝑡 (𝜆𝑡)𝑛
∫0 𝑑𝐺(𝜆)
𝑛!
𝑃[Λ ≤ 𝑥|𝑁(𝑡) = 𝑛] = ∞𝑒−𝜆𝑡 (𝜆𝑡)𝑛
.
∫0 𝑑𝐺(𝜆)
𝑛!

3
 2. Renewal process

Definition: A renewal process is a sequence of i.i.d. random variables {𝑋𝑖 }1∞ with distribution
𝐹(𝑥).
𝑋1 𝑋2 𝑋3 …
𝐹(𝑥) 𝐹(𝑥) 𝐹(𝑥)

Engineering interpretation: An object starts functioning at 𝑡 = 0, fails in accordance with 𝐹(𝑥),

instantaneously repaired to be as good as new state, then the process restarts etc…
It is clear that,
𝑆0 = 0, 𝑆𝑛 = ∑𝑛𝑖=1 𝑋𝑖 , and 𝑁(𝑡) = sup{𝑛: 𝑆𝑛 ≤ 𝑡} … (2.1)

Now, (2.1) is the formal definition of this specific counting process, called the renewal process.

Note: Poisson process is a specific case of the renewal process when 𝐹(𝑥) = 1 − 𝑒 −𝜆𝑥 .

The delayed renewal process:

Only the first cycle is differently distributed in a renewal process.

Distribution of 𝑵(𝒕)
𝑁(𝑡) ≥ 𝑛 ↔ 𝑆𝑛 ≤ 𝑡
Therefore,
𝑃[𝑁(𝑡) = 𝑛] = 𝑃[𝑁(𝑡) ≥ 𝑛] − 𝑃[𝑁(𝑡) ≥ 𝑛 + 1]
(𝑃[𝑁(𝑡) ≥ 𝑛] is the survival function for 𝑁(𝑡), and therefore is a decreasing function.)
= 𝑃[𝑆𝑛 ≤ 𝑡] − 𝑃[𝑆𝑛+1 ≤ 𝑡] …(*)

𝑆𝑛 = ∑𝑛𝑖=1 𝑋𝑖 is distributed as the 𝑛-fold convolution of 𝐹(𝑥) with itself. Denote it by 𝐹 𝑛 (𝑥).
So, from (*), we can write,
𝑃[𝑁(𝑡) = 𝑛] = 𝐹 𝑛 (𝑡) − 𝐹 𝑛+1 (𝑡)

4
 Reminder:
Convolution:
For two random variable 𝑿, 𝒀 distributed as 𝑭(𝒙) and 𝑮(𝒙) respectively. Then the
density of (𝑿 + 𝒀) is given by convolution
𝒕 𝒕
𝒅(𝑭(𝒕) ∗ 𝑮(𝒕)) = ∫𝟎 𝒇(𝒙)𝒈(𝒕 − 𝒙) 𝒅𝒙 = ∫𝟎 𝒈(𝒙)𝒇(𝒕 − 𝒙) 𝒅𝒙

Definition:
The function 𝑚(𝑡) = 𝐸[𝑁(𝑡)] showing the expected number of renewals in [0, 𝑡) is called the
renewal function.

Theorem 4: 𝒎(𝒕) = ∑∞ 𝒏
𝒏=𝟏 𝓕 (𝒕).

Proof:
𝑁(𝑡) = ∑∞
𝑛=1 𝐼𝑛 ,

where 𝐼𝑛 is the indicator function given by 𝐼𝑛 = 1 for 𝑛𝑡ℎ renewal occurrence in [0, 𝑡) and 𝐼𝑛 =
0 otherwise.
Then,

𝐸(𝑁(𝑡)) = 𝐸[∑∞ ∞ ∞
𝑛=1 𝐼𝑛 ] = ∑𝑛=1 𝐸(𝐼𝑛 ) = ∑𝑛=1 𝑃[𝐼𝑛 = 1]

= ∑∞ ∞ 𝑛
𝑛=1 𝑃[𝑆𝑛 ≤ 𝑡] = ∑𝑛=1 ℱ (𝑡)

Examples of some nice convolution:

𝑡−𝜇
1. Normal distribution: 𝐹(𝑡) = Φ ( ), where Φ(. ) is the standard normal c.d.f.
𝜎

The sum of independently normally distributed random variables is again normally distributed.
𝑡−𝑛𝜇
This is given by 𝐹 (𝑛) (𝑡) = 𝑃[𝑁(𝑡) ≥ 𝑛] = Φ ( 𝜎 ).
√𝑛

𝑡−𝑛𝜇 𝑡−(𝑛+1)𝜇
𝑃[𝑁(𝑡) = 𝑛] = Φ ( 𝜎 ) − Φ( ) , 𝑛 = 1,2, …
√𝑛 𝜎√𝑛+1

2. Erlang distribution: This is a specific case of gamma when 𝑛 is integer. The sum of 𝑛
exponential random variables with parameter 𝜆.
(𝜆𝑡)𝑖
Then, 𝐹 (𝑛) (𝑡) = 𝑒 −𝜆𝑡 ∑∞
𝑖=𝑛 .
𝑖!

5
Renewal equations:
Theorem 5:

𝑚(𝑡) = 𝐸(𝑁(𝑡)) = ∑∞
𝑛=1 𝑃[𝑁(𝑡) ≥ 𝑛]

Proof:

𝐸(𝑁(𝑡)) = ∑∞
𝑛=1 𝑛𝑃[𝑁(𝑡) = 𝑛]

= 𝑃[𝑁(𝑡) = 1] + 2𝑃[𝑁(𝑡) = 2] + 3𝑃[𝑁(𝑡) = 3] + ⋯

= (𝑃[𝑁(𝑡) = 1] + 𝑃[𝑁(𝑡) = 2] + 𝑃[𝑁(𝑡) = 3] + ⋯) + (𝑃[𝑁(𝑡) = 2] + 𝑃[𝑁(𝑡) = 3] +
𝑃[𝑁(𝑡) = 4] + ⋯) + (𝑃[𝑁(𝑡) = 3] + 𝑃[𝑁(𝑡) = 4] + 𝑃[𝑁(𝑡) = 5] + ⋯) + …
∞
= ∑∞ ∞
𝑛=1 ∑𝑘=𝑛 𝑃[𝑁(𝑡) = 𝑘] = ∑𝑛=1 𝑃[𝑁(𝑡) ≥ 𝑛]

(𝑛) 𝑡
(𝑛)
𝑚(𝑡) = ∑∞
𝑛=1 𝐹 (𝑡) = 𝐹(𝑡) + ∑∞
𝑛=1 ∫0 𝐹 (𝑡 − 𝑥) 𝑓(𝑥)𝑑𝑥
𝑡 (𝑛) 𝑡
= 𝐹(𝑡) + ∫0 (∑∞
𝑛=1 𝐹 (𝑡 − 𝑥)) 𝑓(𝑥)𝑑𝑥 = 𝐹(𝑡) + ∫0 𝑚(𝑡 − 𝑥) 𝑓(𝑥)𝑑𝑥

Thus, we have the famous Renewal Equation:

𝑡
𝑚(𝑡) = 𝐹(𝑡) + ∫0 𝑚(𝑡 − 𝑥) 𝑓(𝑥)𝑑𝑥

Definition:
The renewal density function is given by
𝑑𝑚(𝑡)
𝑚′ (𝑡) =
𝑑𝑡
Note that this is the rate of corresponding point process.

Differentiating both sides of the renewal equation with respect to 𝑡, we get

𝑡
𝑚′ (𝑡) = 𝐹(𝑡) + ∫0 𝑚′ (𝑡 − 𝑥) 𝑓(𝑥)𝑑𝑥

It can be proved that these equations have unique solutions. We shall prove it with an integral
transformation method called the Laplace transformation.

6
 Laplace transformation

Definition:
If the following conditions satisfy,
i. 𝑓(𝑥) is a real valued function
ii. 𝑓(𝑥) is piecewise continuous in [0, ∞)
iii. There exists constants 𝑎 and 𝑏 such that 𝑓(𝑥) ≤ 𝑏𝑒 𝑎𝑥 , 𝑥 ≥ 0, 𝑎 > 0 (It means that
it can increase not faster than the exponential function).
Then,
∞
𝑓(𝑠) = ∫0 𝑒 −𝑠𝑥 𝑓(𝑥) 𝑑𝑥 is called the Laplace transformation of 𝑓(𝑥).

For simplicity, let us assume that 𝑠 is a real number. If 𝑓(𝑥) is the density of a positive random
variable, then we have a simple probabilistic interpretation that 𝑓(𝑠) = 𝐸[𝑒 −𝑠𝑥 ], where 𝑋 is a
random variable with density 𝑓(𝑥).
Notation: 𝑓(𝑠) = 𝐿(𝑓(𝑡))

Properties:

i. 𝐿(𝑓1 (𝑡) + 𝑓2 (𝑡)) = 𝐿(𝑓1 (𝑡)) + 𝐿(𝑓2 (𝑡))

ii. 𝐿(𝛼𝑓(𝑡)) = 𝛼𝐿(𝑓(𝑡))

Example: Suppose that 𝑓(𝑥) = 𝑒 𝛼𝑥 .

∞ 1
Then 𝐿(𝑒 𝛼𝑥 ) = ∫0 𝑒 −𝑠𝑥 𝑒 𝛼𝑥 𝑑𝑥 = 𝑠−𝛼 , 𝑠 > 𝛼.

iii. 𝐿(𝑓(𝑡 − 𝛼)) = 𝑒 −𝛼𝑡 𝐿(𝑓(𝑡))

iv. 𝐿(𝑓 ′ (𝑡)) = 𝑠𝐿(𝑓(𝑡)) − 𝑓(0) or more generally,
𝐿(𝑓 𝑛 (𝑡)) = 𝑠 𝑛 𝐿(𝑓(𝑡)) − 𝑠 𝑛−1 𝑓(0) − 𝑠 𝑛−2 𝑓 ′ (0) − ⋯ − 𝑓 𝑛−1 (0)
where 𝑓 𝑛 (𝑡) is the 𝑛𝑡ℎ order derivative of 𝑓(𝑡).
𝑡 𝐿(𝑓(𝑡))
v. 𝐿 (∫0 𝑓(𝑢) 𝑑𝑢) = 𝑠
vi. The following formula is of special importance. The Laplace transformation of a
convolution (𝑓1 ∗ 𝑓2 ):
𝑡
𝐿(𝑓1 ∗ 𝑓2 (𝑡)) = 𝐿 (∫0 𝑓1 (𝑢)𝑓2 (𝑡 − 𝑢) 𝑑𝑢) = 𝐿(𝑓1 (𝑡)) 𝐿(𝑓2 (𝑡))

7
 𝑡
Theorem 6: 𝐿(𝑓1 ∗ 𝑓2 (𝑡)) = 𝐿 (∫0 𝑓1 (𝑢)𝑓2 (𝑡 − 𝑢) 𝑑𝑢) = 𝐿(𝑓1 (𝑡)) 𝐿(𝑓2 (𝑡))

Proof:
∞ 𝑡
𝐿(𝑓1 ∗ 𝑓2 (𝑡)) = ∫0 𝑒 −𝑠𝑡 ∫0 𝑓2 (𝑡 − 𝑢)𝑓1 (𝑢) 𝑑𝑢
∞ ∞
= ∫0 𝑒 −𝑠𝑢 𝑓1 (𝑢) ∫𝑢 𝑒 −𝑠(𝑡−𝑢) 𝑓2 (𝑡 − 𝑢) 𝑑𝑡 𝑑𝑢
∞ ∞
= ∫0 𝑒 −𝑠𝑢 𝑓1 (𝑢) ∫0 𝑒 −𝑠𝑦 𝑓2 (𝑦) 𝑑𝑦 𝑑𝑢

= 𝐿(𝑓1 (𝑡)) 𝐿(𝑓2 (𝑡))

Here we have used the Dirichlet’s formula given by,

𝑧 𝑦 𝑧 𝑧
∫0 ∫0 𝑓(𝑥, 𝑦) 𝑑𝑥 𝑑𝑦 = ∫0 ∫𝑥 𝑓(𝑥, 𝑦) 𝑑𝑦 𝑑𝑥.

Next, we provide a table for Laplace transformation for some functions.

𝒇(𝒕) 𝑳(𝒇(𝒕))
1 1/𝑠
𝑡 1/𝑠 2
𝑡𝑛 𝑛+1
𝑛!/𝑠 , 𝑛 = 1,2,3, …
𝑡𝛼 Γ(𝛼 + 1)/𝑠 𝑛+1
𝑒 𝛼𝑡 1/(𝑠 − 𝛼), 𝑠 > 𝛼
𝑒 𝛼𝑡 𝑡 𝑛 𝑛!/ (𝑠 − 𝛼)𝑛 , 𝑠 > 𝛼
Sin(𝑤𝑡) 𝑤/(𝑠 2 + 𝑤 2 )
Cos(𝑤𝑡) 𝑠/(𝑠 2 + 𝑤 2 )

Now we are able to solve the renewal equation in terms of Laplace transformation applying the
convolution property from property (vi), i.e., Theorem 6.
Therefore, the renewal density function is written as

𝐿(𝑚′ (𝑡)) = 𝐿(𝑓(𝑡)) + 𝐿(𝑚′ (𝑡))𝐿(𝑓(𝑡))

𝐿(𝑓(𝑡))
 𝐿(𝑚′ (𝑡)) = 1−𝐿(𝑓(𝑡)) …(a)

Applying Laplace transformation property (v) to Eq. (a) above, we get

𝐿(𝑓(𝑡))
𝐿(𝑚(𝑡)) = . …(b)
𝑠(1−𝐿(𝑓(𝑡)))

In order to obtain 𝑚(𝑡), we must make an inverse transformation. It can be done explicitly for
simple specific cases.

8
Example 1: Let us consider the exponential distribution which is the simplest case.

𝜆
𝑓(𝑡) = 𝜆𝑒 −𝜆𝑡 , 𝐿(𝑓(𝑡)) = 𝑠+𝜆
Therefore,
𝜆
( ) 𝜆
𝑠+𝜆
𝐿(𝑚(𝑡)) = 𝜆 = 𝑠2
𝑠(1− )
𝑠+𝜆
From the table in Page 8,
𝑚(𝑡) = 𝜆𝑡, 𝑚′ (𝑡) = 𝜆
𝑡
Obtain the solution of 𝑚(𝑡) = 𝑓(𝑡) + ∫0 𝑓(𝑥)𝑚(𝑡 − 𝑥) 𝑑𝑥 using 𝐿 transformation and a
specific one for 𝑓(𝑡) = 1 −𝑒 −𝜆𝑡 . Note that 𝐿(𝑡) = 1/𝑠 2 . Then,
𝐿(𝑚(𝑡)) = 𝐿(𝑓(𝑡)) + 𝐿(𝑓(𝑡))𝐿(𝑚(𝑡))
𝐿(𝐹(𝑡)) 𝐿(𝑓(𝑡))
 𝐿(𝑚(𝑡)) = 1−𝐿(𝑓(𝑡)) = .
𝑠(1−𝐿(𝑓(𝑡)))
We know that,
𝑓(𝑡) = 𝜆𝑒 −𝜆𝑡 .
∞
Therefore, 𝐿(𝑓(𝑡)) = ∫0 𝑒 −𝑠𝑥 𝜆𝑒 −𝜆𝑥 𝑑𝑥 = 𝜆/(𝜆 + 𝑠) .
𝜆
So, we get, 𝐿(𝑚(𝑡)) = 𝑠2 and from the table in Page 8, we get 𝑚(𝑡) = 𝜆𝑡

Example 2: Let us now consider the Erlang distribution. The density function is given by

𝜆𝑚 𝑡 𝑚−1 𝑒 −𝜆𝑡
𝑓(𝑡) = (𝑚−1)!
, 𝑚 = 0, 1, 2, …

The distribution function is given by,

(𝜆𝑡)𝑖
𝐹(𝑡) = 1 −𝑒 −𝜆𝑡 ∑𝑚−1
𝑖=0 𝑖!

∞ (𝜆𝑡)𝑖
It can be shown that, 𝑚(𝑡) = 𝑒 −𝜆𝑡 ∑∞
𝑛=1 ∑𝑖=𝑚𝑛 𝑖!

1 1 1
Note that, for 𝑚 = 1, 𝑚(𝑡) = 𝜆𝑡, and for 𝑚 = 2, 𝑚(𝑡) = 2 [𝜆𝑡 − 2 + 2 𝑒 −2𝜆𝑡 ]

9
 Example 3:
Let us now consider that the cycle length is normally distributed with mean 𝜇 and 𝜎, where
𝜇 > 3𝜎. We consider 𝜇 > 3𝜎 to practically ensure positive values, though the theoretical
range goes from −∞ to ∞.
Then,
(𝑛) 𝑡−𝑛𝜇
𝑚(𝑡) = ∑∞𝑛=1 𝐹 (𝑡) ≈ ∑∞
𝑛=1 Φ ( 𝜎 𝑛 )
√

3. Recurrence times

Denote by {𝑅(𝑡), 𝑡 ≥ 0}, the stochastic process of backward recurrence times,

𝑅(𝑡) = 𝑡 − 𝑆𝑁(𝑡) , where 𝑆𝑁(𝑡) is the time of the last renewal before time 𝑡.

Denote by {𝑉(𝑡), 𝑡 ≥ 0}, the stochastic process of forward recurrence times,

𝑉(𝑡) = 𝑆𝑁(𝑡)+1 − 𝑡 , where 𝑆𝑁(𝑡)+1 is the time of the first renewal after time 𝑡.

Next, denote by 𝐹𝑅(𝑡) (𝑥) = 𝑃[𝑅(𝑡) ≤ 𝑥] and 𝐹𝑉(𝑡) (𝑥) = 𝑃[𝑉(𝑡) ≤ 𝑥], the distribution
functions of the backward and forward renewal times, respectively.

Theorem 7: Derive the distribution functions of 𝑅(𝑡) and 𝑉(𝑡).

Proof.

𝐹𝑅(𝑡) (𝑥) = 𝑃[𝑡 − 𝑆𝑁(𝑡) ≤ 𝑥] = 𝑃[𝑆𝑁(𝑡) ≥ 𝑡 − 𝑥]

= ∑∞
𝑛=1 𝑃[𝑡 − 𝑥 ≤ 𝑆𝑛 , 𝑁(𝑡) = 𝑛]

= ∑∞
𝑛=1 𝑃[𝑡 − 𝑥 ≤ 𝑆𝑛 ≤ 𝑡 ≤ 𝑆𝑛+1 ]
𝑡
= ∑∞ ̅ 𝑛 𝑛
𝑛=1 ∫(𝑡−𝑥) 𝐹 (𝑡 − 𝑢) 𝑑𝐹 (𝑢) [𝐹 (𝑢) is the 𝑛-fold convolution of 𝐹(𝑢)]

𝑡
= ∫(𝑡−𝑥) 𝐹̅ (𝑡 − 𝑢) 𝑑(∑∞ 𝑛
𝑛=1 𝐹 (𝑢))

𝑡 𝑡
= ∫𝑡−𝑥 𝐹̅ (𝑡 − 𝑢) 𝑑𝑚(𝑢) = ∫𝑡−𝑥 𝐹̅ (𝑡 − 𝑢) 𝑚′ (𝑢)𝑑𝑢

Therefore,

10
 𝑡
∫ 𝐹̅ (𝑡 − 𝑢)𝑚′ (𝑢)𝑑𝑢 , 0 ≤ 𝑥 ≤ 𝑡
𝐹𝑅(𝑡) (𝑥) = { 𝑡−𝑥
1,𝑥 > 𝑡

and the density is given by,

𝐹̅ (𝑥)𝑚′ (𝑡 − 𝑥) , 0 ≤ 𝑥 ≤ 𝑡
𝑓𝑅(𝑡) (𝑥) = {
0 ,𝑥 > 𝑡

Now, for the forward recurrence time:

𝐹𝑉(𝑡) (𝑥) = 𝑃[𝑆𝑁(𝑡)+1 − 𝑡 ≤ 𝑥] = 𝑃[𝑆𝑁(𝑡)+1 ≤ 𝑥 + 𝑡]

= ∑∞
𝑛=0 𝑃[𝑆𝑛+1 ≤ 𝑥 + 𝑡, 𝑁(𝑡) = 𝑛]
∞
= 𝐹(𝑡 + 𝑥) − 𝐹(𝑡) + ∑∞
𝑛=1 ∫0 [𝐹(𝑡 + 𝑥 − 𝑢) − 𝐹(𝑡 − 𝑢)] 𝑑𝐹𝑆𝑛 (𝑢)

∞
= 𝐹(𝑡 + 𝑥) − 𝐹(𝑡) + ∫0 [𝐹(𝑡 + 𝑥 − 𝑢) − 𝐹(𝑡 − 𝑢)] 𝑚′ (𝑢) 𝑑𝑢
𝑡
(From the main renewal equation: 𝐹(𝑡) = 𝑚(𝑡) − ∫0 𝐹(𝑡 − 𝑢) 𝑚′ (𝑢)𝑑𝑢
𝑡 𝑡
where the convolution property: ∫0 𝑚(𝑡 − 𝑥)𝑓(𝑥) 𝑑𝑥 = ∫0 𝐹(𝑡 − 𝑥)𝑚′ (𝑥) 𝑑𝑥.)

We get,
𝑡
𝐹𝑉(𝑡) (𝑥) = 𝐹(𝑡 + 𝑥) − ∫0 𝐹̅ (𝑡 + 𝑥 − 𝑢)𝑚′(𝑢) 𝑑𝑢

and the density is given by

𝑡
𝑓𝑉(𝑡) (𝑥) = 𝑓(𝑡 + 𝑥) − ∫0 𝑓(𝑡 + 𝑥 − 𝑢)𝑚′(𝑢) 𝑑𝑢

Hence the theorem is proved.

Limiting behaviour:
Let 𝑁(∞) = lim 𝑁(𝑡) be the total number of renewals in [0,∞).
𝑡→∞

It is clear that 𝑁(∞) → ∞ with probability 1, since the only way 𝑁(∞) can be finite is for one
of the interarrival times to be infinite.

11
 𝑁(𝑡) 1 𝑡
Theorem 8: With probability 1, →𝑡→∞ , and 𝑁(𝑡) →𝑡→∞ 𝜇 .
𝑡 𝜇

Proof.
𝑆𝑁(𝑡) 𝑡 𝑆𝑁(𝑡)+1
As 𝑆𝑁(𝑡) ≤ 𝑡 ≤ 𝑆𝑁(𝑡)+1, ≤ 𝑁(𝑡) ≤ .
𝑁(𝑡) 𝑁(𝑡)

By the strong law of large numbers,

𝑆𝑁(𝑡)
→ 𝜇 as 𝑁(𝑡) → ∞ when 𝑡 → ∞.
𝑁(𝑡)

𝑆𝑁(𝑡)+1 𝑆
𝑁(𝑡)+1 𝑁(𝑡)+1
Now, let us write, = (𝑁(𝑡)+1 )( ).
𝑁(𝑡) 𝑁(𝑡)

𝑆
𝑁(𝑡)+1 𝑁(𝑡)+1
Note that, (𝑁(𝑡)+1 ) → 𝜇 and ( ) → 1 as 𝑁(𝑡) → ∞.
𝑁(𝑡)

12